Generalized Kelly Criterion

• The generalized Kelly criterion is a versatile framework that extends the classical log-optimal betting rule to incorporate practical utility functions, variable pay-offs, and extraneous wealth.
• It integrates sophisticated risk controls and temporal, multivariate dependencies by embedding the betting strategy within constrained optimization and probabilistic models.
• Extensions of the criterion span continuous-time processes, portfolio rebalancing, and even quantum settings, highlighting its adaptability to complex financial and stochastic environments.

Searching arXiv for recent and foundational papers on generalized Kelly criterion. arxiv_search(query="generalized Kelly criterion", max_results=10, sort_by="relevance") The generalized Kelly criterion comprises a family of extensions of Kelly’s strategy—described in one paper as “optimal for simple investment games with log utility”—to settings with practical utility functions, extraneous wealth, variable pay-off, temporal correlation, multivariate portfolios, explicit risk controls, and nonclassical state spaces. In the formulation of “Generalizing the Kelly strategy,” the central result is that, for “any continuous, concave, differentiable utility function,” the optimal choice at every point depends only on “the probability of reaching that point,” and that a binomial expansion can reduce the problem size “from exponential to quadratic” (Viswanathan, 2016).

1. Classical benchmark and the scope of generalization

The classical benchmark is the repeated favorable bet with logarithmic utility. In the even-money memoryless coin-flip model, wealth evolves as

$V_{k+1}=(1+KX_k)V_k,$

with $X_k\in\{+1,-1\}$, and the expected log-growth per play is

$\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$

The unique maximizer is

$K^*=2p-1.$

This is the form that “pervades the literature” in the i.i.d. even-money coin-flip case (O'Brien et al., 2020).

A more general Bernoulli bet with gain fraction $b>0$ and loss fraction $a\in(0,1]$ yields

$g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$

with unique maximizer

$f^*=\frac{pb-(1-p)a}{ab}.$

For arbitrary i.i.d. returns $X$, the long-run growth rate becomes

$g(f)=\mathbb E[\ln(1+fX)],$

and strict concavity gives the first-order condition

$X_k\in\{+1,-1\}$0

These expressions provide the reference point from which later generalizations depart (Lototsky et al., 2020).

Within this literature, “generalized Kelly criterion” therefore refers not to a single modification but to a class of formulations in which either the objective, the return process, the admissible strategies, the market structure, or the information set differs from the classical i.i.d. log-utility benchmark. Some generalizations preserve expected logarithmic growth exactly; others embed it in larger utility-theoretic or risk-constrained programs (Viswanathan, 2016).

2. Utility functions, extraneous wealth, and variable pay-off

A direct generalization replaces simple log utility by a “large class of practical utility functions” and includes “the effect of extraneous wealth.” The abstract of “Generalizing the Kelly strategy” states a counterintuitive theorem: for “any continuous, concave, differentiable utility function, the optimal choice at every point depends only on the probability of reaching that point.” The same note states that practical calculation is enabled “through use of the binomial expansion,” reducing the problem size “from exponential to quadratic,” with applications including “(better) automatic investing and risk taking under uncertainty” (Viswanathan, 2016).

A different axis of generalization concerns random pay-off conditional on a win. In “Kelly criterion for variable pay-off,” the gambler stakes a fraction $X_k\in\{+1,-1\}$1, wins with probability $X_k\in\{+1,-1\}$2, loses with probability $X_k\in\{+1,-1\}$3, and on a win receives a nonnegative random pay-off $X_k\in\{+1,-1\}$4 with density or mass function $X_k\in\{+1,-1\}$5. The expected logarithmic growth per round is

$X_k\in\{+1,-1\}$6

and the optimal fraction $X_k\in\{+1,-1\}$7 is the unique solution of the fundamental integral equation

$X_k\in\{+1,-1\}$8

The same paper states that this optimal fraction is “smaller than the classical Kelly fraction for the same game with the constant average pay-off,” with equality only when the pay-off distribution is degenerate (Pérez-Marco, 2014).

Taken together, these results broaden the Kelly program in two distinct ways. The first holds the probabilistic tree fixed while enlarging the utility class and accounting for outside wealth. The second keeps expected log-growth but replaces constant pay-off by a random multiplier, so that the optimal stake is determined by an integral equation rather than a closed-form fraction. This suggests that “generalized Kelly criterion” is as much about preserving concavity and long-run growth logic under richer primitives as it is about producing new formulas.

3. Temporal correlation, side information, and serial dependence

Several generalizations abandon the i.i.d. assumption. In “A Generalization of the Classical Kelly Betting Formula to the Case of Temporal Correlation,” the probability of heads on flip $X_k\in\{+1,-1\}$9 depends on previous flips. For memory depth $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$0,

$\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$1

and the optimal fixed fraction over $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$2 flips is

$\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$3

In the $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$4 closed form,

$\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$5

The paper also states that the time-varying strategy $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$6 does best, and that the best fixed-$\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$7 bettor $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$8 equals the average of these $\mathrm{ELG}(K)=p\log(1+K)+(1-p)\log(1-K).$9 (O'Brien et al., 2020).

Another information-theoretic generalization treats the value of messages or side information. “Pragmatic Information Rates, Generalizations of the Kelly Criterion, and Financial Market Efficiency” defines pragmatic information as the mutual information between actions and messages, and proves existence of a per-symbol pragmatic information rate for jointly stationary processes. In the horse-race formulation, the one-race doubling increment can be written as

$K^*=2p-1.$0

so the optimal choice is $K^*=2p-1.$1. With side information $K^*=2p-1.$2, the gain in optimal doubling rate is

$K^*=2p-1.$3

For a stationary sequence of races, the long-term excess doubling rate from using side information is bounded by the mutual-information rate $K^*=2p-1.$4 (0903.2243).

Serial dependence also appears in market models. “Beating the Best Constant Rebalancing Portfolio in Long-Term Investment” groups returns into $K^*=2p-1.$5-length blocks $K^*=2p-1.$6 and defines $K^*=2p-1.$7-log-optimal portfolios by

$K^*=2p-1.$8

Under a block-wise i.i.d. model, the corresponding $K^*=2p-1.$9-cyclic constant strategy attains the highest sustainable growth rate $b>0$0, and a $b>0$1-parallel universal portfolio learns this dependence online without knowing the joint law (Lam, 8 Jul 2025).

A related but distinct extension treats rebalancing frequency itself as a design variable. “Necessary and Sufficient Conditions for Frequency-Based Kelly Optimal Portfolio” defines

$b>0$2

for $b>0$3-step rebalancing, derives KKT-style necessary and sufficient conditions, and proves the “Extended Dominant Asset Theorem”: asset $b>0$4 is dominant if and only if the Kelly-optimal portfolio is $b>0$5 (Hsieh, 2020).

4. Multivariate portfolios and simultaneous bets

The multivariate generalization replaces a single favorable gamble by a portfolio of assets or simultaneous binary bets. In the general portfolio form of “Generalized framework for applying the Kelly criterion to stock markets,” the investor allocates fractions $b>0$6, holds the remainder in cash, and maximizes

$b>0$7

where $b>0$8 is the vector of excess returns. In the small-bet approximation,

$b>0$9

with $a\in(0,1]$0 and $a\in(0,1]$1. The same paper states that this matrix-inversion framework works “for one or a portfolio of stocks,” and that the Kelly fractions can be calculated for “the Gaussian distribution and correlated multivariate assets” (Byrnes et al., 2018).

A portfolio formulation with an explicit risk-return parameter appears in “Kelly’s Criterion in Portfolio Optimization: A Decoupled Problem.” The coupled return objective is

$a\in(0,1]$2

while the decoupled return function is

$a\in(0,1]$3

The single-objective program is

$a\in(0,1]$4

subject to $a\in(0,1]$5 and box constraints. Here $a\in(0,1]$6 interpolates between pure Kelly-type growth and pure variance minimization (Peterson, 2017).

For simultaneous binary bets, “Efficient Multivariate Kelly Optimization Reveals Sigmoidal Scaling Laws” studies the objective

$a\in(0,1]$7

in a discrete-scenario market, and specializes to $a\in(0,1]$8 independent binary bets. The paper states that naive evaluation costs $a\in(0,1]$9, whereas an integral-transform formulation reduces evaluation of the objective “from $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$0 to $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$1.” It also introduces decomposition-based lower and upper bounds and reports that the shortfall ratio between them is “well-approximated by a sigmoid function of the relative subproblem size,” with per-instance fits having median $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$2 and held-out prediction median $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$3 (Tepelyan et al., 27 Apr 2026).

These formulations show that the generalized Kelly criterion in portfolio settings is not merely a scalar fraction formula. It becomes a concave optimization problem over a simplex or a constrained feasible set, often with state-space size, dependence structure, and numerical tractability as first-order concerns.

5. Risk control, constraints, and robustness

The classical Kelly rule is often described in this literature as maximizing long-run growth but being “risky.” “Phase transitions in optimal strategies for betting” formalizes a mean-fluctuation trade-off using

$g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$4

with $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$5 and $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$6. The stationarity condition is

$g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$7

As $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$8, one recovers the Kelly point $g(f)=p\ln(1+fb)+(1-p)\ln(1-fa),$9; as $f^*=\frac{pb-(1-p)a}{ab}.$0, one obtains the risk-free or null strategy $f^*=\frac{pb-(1-p)a}{ab}.$1. The paper identifies a first-order transition at a critical $f^*=\frac{pb-(1-p)a}{ab}.$2, proves a thermodynamic-uncertainty-relation-type bound

$f^*=\frac{pb-(1-p)a}{ab}.$3

and states that the framework gives a one-parameter family of “fractional-Kelly” strategies (Dinis et al., 2020).

Constraint-based generalizations modify the feasible set directly. In “Sizing the bets in a focused portfolio,” the long-run growth objective is

$f^*=\frac{pb-(1-p)a}{ab}.$4

subject to long-only constraints $f^*=\frac{pb-(1-p)a}{ab}.$5, leverage limit $f^*=\frac{pb-(1-p)a}{ab}.$6, maximum individual allocation $f^*=\frac{pb-(1-p)a}{ab}.$7, and a maximum permanent-loss constraint based on $f^*=\frac{pb-(1-p)a}{ab}.$8. The KKT stationarity equations show explicitly how each active multiplier clips or tilts the unconstrained solution toward the constrained region (Vukcevic et al., 2024).

Robustness to parameter uncertainty motivates yet another generalization. “Tackling estimation risk in Kelly investing using options” augments the one-period binomial Kelly problem by allowing a fraction $f^*=\frac{pb-(1-p)a}{ab}.$9 in a European put option and $X$0 in a bond, with objective

$X$1

For fixed hedge parameter $X$2, the optimal option weight $X$3 has a closed form, and Proposition 2.1 states that the KO-optimal relative payoff coincides almost surely with the classical Kelly payoff at the true parameters. Under misspecification, however, different hedge choices dominate in different regions, and Theorem 3.1 shows that a convex mixture of two such KO strategies asymptotically inherits the better growth rate, thereby “eliminating estimation risk in the long run” (Lillo et al., 26 Aug 2025).

Across these formulations, the generalized Kelly criterion is no longer only about maximizing expected logarithmic growth under perfect-model assumptions. It also includes explicit fluctuation penalties, admissibility constraints, and instruments designed to reduce sensitivity to estimation error.

6. Continuous-time and nonclassical extensions

One direction of generalization changes the time scale and the return process. “Kelly Criterion: From a Simple Random Walk to Lévy Processes” moves from discrete i.i.d. bets to continuous rebalancing under a Lévy process $X$4. If wealth follows

$X$5

then the almost-sure long-run growth rate is

$X$6

and the unique maximizer solves

$X$7

In the geometric Brownian motion case $X$8, this reduces to

$X$9

This is the continuous-time log-optimal or Merton form (Lototsky et al., 2020).

A more radical generalization is quantum. “A quantum double-or-nothing game: The Kelly Criterion for Spins” considers a sequence of spin-$g(f)=\mathbb E[\ln(1+fX)],$0 particles measured in freely chosen polarization directions. If the gambler allocates fractions $g(f)=\mathbb E[\ln(1+fX)],$1 with $g(f)=\mathbb E[\ln(1+fX)],$2, and measurement probabilities are

$g(f)=\mathbb E[\ln(1+fX)],$3

then the expected logarithmic growth per round is

$g(f)=\mathbb E[\ln(1+fX)],$4

and the unique optimum is

$g(f)=\mathbb E[\ln(1+fX)],$5

The paper states that this has “precisely the same form as the classical Kelly rule $g(f)=\mathbb E[\ln(1+fX)],$6,” but that noncommuting measurements create a trade-off between instantaneous growth and information gain. Numerical dynamic programming shows that the quantum strategy differs from the classical strategy and can outperform any static classical Kelly approach “by up to a few percent in long-term growth rate” for non-orthogonal pure-state mixtures (Meister et al., 2023).

These extensions preserve the defining Kelly logic—maximize a long-run logarithmic criterion—while relocating it into continuous-time semimartingale finance or a quantum measurement-and-betting problem. A plausible implication is that the generalized Kelly criterion is best understood as a methodology for growth-optimal allocation under increasingly rich probabilistic structures, rather than as a single closed-form betting rule.